Real-World Validation: 79.6% on UCI HAR with Zero Feature Engineering

From Simulations to Real Data

Our previous benchmark showed streaming streaming neurons outperforming traditional approaches on simulated industrial data. The natural question: does this advantage hold on real-world sensor data?

We tested on the UCI HAR dataset (Anguita et al., 2013) — one of the most widely cited benchmarks in activity recognition research. Real accelerometer and gyroscope data from 30 volunteers performing 6 activities with a smartphone strapped to their waist.

The Scenario: Continuous Activity Monitoring

Standard UCI HAR evaluates window classification: here is a 2.5-second clip, what activity is it? But that is not how wearables actually work. A smartwatch receives a continuous stream of sensor data. The person walks, then sits down, then stands up, then climbs stairs. The system must recognize each activity in real-time as it happens.

We built multi-activity sequences by concatenating real UCI HAR windows from different activities — creating realistic scenarios with 4-6 activity transitions per sequence. This is the continuous monitoring problem that streaming neurons were designed for.

The Activities

• WALKING — periodic acceleration pattern
• WALKING UPSTAIRS — similar to walking but with forward lean
• WALKING DOWNSTAIRS — similar but with different impact pattern
• SITTING — minimal movement, gravity-dominated
• STANDING — very similar to sitting from a sensor perspective
• LAYING — gravity axis shift

The hard pairs: sitting vs standing (nearly identical instantaneous readings) and walking upstairs vs downstairs (subtle differences in impact timing).

Three Approaches Compared

Luviner Streaming: Trained with train_sequential() on multi-activity sequences. At inference, raw 9-axis sensor data is fed one sample at a time. Neuron state accumulates context across activity transitions.

Stateless: Same network, but each sample classified independently. No memory of previous readings.

Windowed Features: Buffer 32 readings, extract 36 statistical features (mean, std, range, roughness per channel). The classical embedded ML approach.

Results

Key Findings

• +11.0% over stateless — streaming memory provides real value on real sensor data
• +49.0% over windowed — statistical features fail to capture activity transitions
• 3x less RAM than the windowed approach, with dramatically better accuracy

Why Windowed Features Collapse

The windowed approach works well for classifying isolated windows (the standard UCI HAR task). But in continuous monitoring with activity transitions, the sliding window contains samples from two different activities during transitions. The mean and standard deviation of walking + sitting readings do not represent either activity well.

Luviner's streaming neurons handle this naturally: their internal state evolves smoothly across transitions, detecting the change from one activity pattern to another.

What This Means

This is the first validation of Luviner on a real public dataset with real sensor data from real humans. The results confirm what we demonstrated on synthetic data:

1. Streaming memory has real value — +11.0% accuracy with no engineering overhead
2. No feature engineering needed — raw 9-axis sensor data, no buffering, no manual statistics
3. Handles transitions — continuous monitoring is where streaming neurons shine

For wearable devices, industrial monitors, and medical sensors that operate in continuous mode, this is the architecture that makes sense.

Try It Yourself

docker exec luviner-edge-ai-1 python -m cli.main har

The benchmark automatically downloads the UCI HAR dataset and runs the full comparison.

Get started with Luviner →